384 lines
12 KiB
Plaintext
384 lines
12 KiB
Plaintext
== Configuration and Integration
|
|
|
|
=== Hazard3 Source Files
|
|
|
|
Hazard3's source is written in Verilog 2005, and is self-contained. It can be found here: https://github.com/Wren6991/Hazard3/tree/master/hdl[github.com/Wren6991/Hazard3/blob/master/hdl]. The file https://github.com/Wren6991/Hazard3/blob/master/hdl/hazard3.f[hdl/hazard3.f] is a list of all the source files required to instantiate Hazard3.
|
|
|
|
Files ending with `.vh` are preprocessor include files used by the Hazard3 source. Two to take note of are:
|
|
|
|
* https://github.com/Wren6991/Hazard3/blob/master/hdl/hazard3_config.vh[hazard3_config.vh]: the main Hazard3 configuration header. Lists and describes Hazard3's global configuration parameters, such as ISA extension support
|
|
* https://github.com/Wren6991/Hazard3/blob/master/hdl/hazard3_config_inst.vh[hazard3_config_inst.vh]: a file which propagates configuration parameters through module instantiations, all the way down from Hazard3's top-level modules through the internals
|
|
|
|
Therefore there are two ways to configure Hazard3:
|
|
|
|
* Directly edit the parameter defaults in `hazard3_config.vh` in your local Hazard3 checkout (and then let the top-level parameters default when instantiating Hazard3)
|
|
* Set all configuration parameters in your Hazard3 instantiation, and let the parameters propagate down through the hierarchy
|
|
|
|
=== Top-level Modules
|
|
|
|
Hazard3 has two top-level modules:
|
|
|
|
* https://github.com/Wren6991/Hazard3/blob/master/hdl/hazard3_cpu_1port.v[hazard3_cpu_1port]
|
|
* https://github.com/Wren6991/Hazard3/blob/master/hdl/hazard3_cpu_2port.v[hazard3_cpu_2port]
|
|
|
|
These are both thin wrappers around the https://github.com/Wren6991/Hazard3/blob/master/hdl/hazard3_core.v[hazard3_core] module. `hazard3_cpu_1port` has a single AHB5 bus port which is shared for instruction fetch, loads, stores and AMOs. `hazard3_cpu_2port` has two AHB5 bus ports, one for instruction fetch, and the other for loads, stores and AMOs. The 2-port wrapper has higher potential for performance, but the 1-port wrapper may be simpler to integrate, since there is no need to arbitrate multiple bus masters externally.
|
|
|
|
The core module `hazard3_core` can also be instantiated directly, which may be more efficient if support for some other bus standard is desired. However, the interface of `hazard3_core` will not be documented and is not guaranteed to be stable.
|
|
|
|
=== Configuration Parameters
|
|
|
|
==== Reset state configuration
|
|
|
|
===== RESET_VECTOR
|
|
|
|
Address of the first instruction executed after Hazard3 comes out of reset.
|
|
|
|
Default value: all-zeroes.
|
|
|
|
===== MTVEC_INIT
|
|
|
|
Initial value of the machine trap vector base CSR (<<reg-mtvec>>).
|
|
|
|
Bits clear in <<param-MTVEC_WMASK>> will never change from this initial value.
|
|
Bits set in <<param-MTVEC_WMASK>> can be written/set/cleared as normal.
|
|
|
|
Default value: all-zeroes.
|
|
|
|
==== RISC-V ISA support
|
|
|
|
[[param-EXTENSION_A]]
|
|
===== EXTENSION_A
|
|
|
|
Support for the A extension: atomic read/modify/write. 0 for disable, 1 for enable.
|
|
|
|
Default value: 1
|
|
|
|
[[param-EXTENSION_C]]
|
|
===== EXTENSION_C
|
|
|
|
Support for the C extension: compressed (variable-width). 0 for disable, 1 for enable.
|
|
|
|
Default value: 1
|
|
|
|
[[param-EXTENSION_M]]
|
|
===== EXTENSION_M
|
|
|
|
Support for the M extension: hardware multiply/divide/modulo. 0 for disable, 1 for enable.
|
|
|
|
Default value: 1
|
|
|
|
[[param-EXTENSION_ZBA]]
|
|
===== EXTENSION_ZBA
|
|
|
|
Support for Zba address generation instructions. 0 for disable, 1 for enable.
|
|
|
|
Default value: 1
|
|
|
|
[[param-EXTENSION_ZBB]]
|
|
===== EXTENSION_ZBB
|
|
|
|
Support for Zbb basic bit manipulation instructions. 0 for disable, 1 for enable.
|
|
|
|
Default value: 1
|
|
|
|
[[param-EXTENSION_ZBC]]
|
|
===== EXTENSION_ZBC
|
|
|
|
Support for Zbc carry-less multiplication instructions. 0 for disable, 1 for enable.
|
|
|
|
Default value: 1
|
|
|
|
[[param-EXTENSION_ZBS]]
|
|
===== EXTENSION_ZBS
|
|
|
|
Support for Zbs single-bit manipulation instructions. 0 for disable, 1 for enable.
|
|
|
|
Default value: 1
|
|
|
|
[[param-EXTENSION_ZBKB]]
|
|
===== EXTENSION_ZBKB
|
|
|
|
Support for Zbkb basic bit manipulation for cryptography.
|
|
|
|
Requires: <<param-EXTENSION_ZBB>>. (Since Zbb and Zbkb have a large overlap, this flag enables only those instructions which are in Zbkb but aren't in Zbb. Therefore both flags must be set for full Zbkb support.)
|
|
|
|
Default value: 1
|
|
|
|
[[param-EXTENSION_XH3B]]
|
|
===== EXTENSION_XH3B
|
|
|
|
Custom bit manipulation instructions for Hazard3: `h3.bextm` and `h3.bextmi`.
|
|
|
|
Default value: 1
|
|
|
|
[[param-EXTENSION_ZIFENCEI]]
|
|
===== EXTENSION_ZIFENCEI
|
|
|
|
Support for the fence.i instruction. When the branch predictor is not present,
|
|
this instruction is optional, since a plain branch/jump is sufficient to
|
|
flush the instruction prefetch queue. When the branch predictor is enabled
|
|
(<<param-BRANCH_PREDICTOR>> is 1), this instruction must be implemented.
|
|
|
|
Default value: 1
|
|
|
|
==== CSR support
|
|
|
|
NOTE: the Zicsr extension is implied by any of <<param-CSR_M_MANDATORY>>, <<param-CSR_M_TRAP>>,
|
|
<<param-CSR_COUNTER>>.
|
|
|
|
[[param-CSR_M_MANDATORY]]
|
|
===== CSR_M_MANDATORY
|
|
|
|
Bare minimum CSR support e.g. <<reg-misa>>. This flag is an absolute
|
|
requirement for compliance with the RISC-V privileged specification. However,
|
|
the privileged specification itself is an optional extension. Hazard3 allows
|
|
the mandatory CSRs to be disabled to save a small amount of area in
|
|
deeply-embedded implementations.
|
|
|
|
[[param-CSR_M_TRAP]]
|
|
===== CSR_M_TRAP
|
|
|
|
Include M-mode trap-handling CSRs, and enable trap support.
|
|
|
|
[[param-CSR_COUNTER]]
|
|
===== CSR_COUNTER
|
|
|
|
Include the basic performance counters (`cycle`/`instret`) and relevant CSRs. Note that these performance counters are now in their own separate extension (Zicntr) and are no longer mandatory.
|
|
|
|
[[param-U_MODE]]
|
|
===== U_MODE
|
|
|
|
Support the U (user) privilege level. In U-mode, the core performs unprivileged
|
|
bus accesses, and software's access to CSRs is restricted. Additionally, if
|
|
the PMP is included, the core may restrict U-mode software's access to
|
|
memory.
|
|
|
|
Requires: <<param-CSR_M_TRAP>>.
|
|
|
|
[[param-PMP_REGIONS]]
|
|
===== PMP_REGIONS
|
|
|
|
Number of physical memory protection regions, or 0 for no PMP. PMP is more
|
|
useful if U mode is supported, but this is not a requirement.
|
|
|
|
Hazard3's PMP supports only the NAPOT and(if <<param-PMP_GRAIN>> is 0) NA4
|
|
region types.
|
|
|
|
Requires: <<param-CSR_M_TRAP>>.
|
|
|
|
[[param-PMP_GRAIN]]
|
|
===== PMP_GRAIN
|
|
|
|
This is the _G_ parameter in the privileged spec, which defines the
|
|
granularity of PMP regions. Minimum PMP region size is 1 <<(_G_ + 2) bytes.
|
|
|
|
If _G_ > 0, `pmcfg.a` can not be set to NA4 (attempting to do so will set the
|
|
region to OFF instead).
|
|
|
|
If _G_ > 1, the _G_ - 1 LSBs of pmpaddr are read-only-0 when `pmpcfg.a` is
|
|
OFF, and read-only-1 when `pmpcfg.a` is NAPOT.
|
|
|
|
Default value: 0
|
|
|
|
[[param-PMP_HARDWIRED]]
|
|
===== PMP_HARDWIRED
|
|
|
|
PMPADDR_HARDWIRED: If a bit is 1, the corresponding region's pmpaddr and
|
|
pmpcfg registers are read-only, with their values fixed when the processor is
|
|
instantiated. PMP_GRAIN is ignored on hardwired regions.
|
|
|
|
Hardwired regions are far cheaper, both in area and comparison delay, than
|
|
dynamically configurable regions.
|
|
|
|
Hardwired PMP regions are a good option for setting default U-mode permissions
|
|
on regions which have access controls outside of the processor, such as
|
|
peripheral regions. For this case it's recommended to make hardwired regions
|
|
the highest-numbered, so they can be overridden by lower-numbered dynamic
|
|
regions.
|
|
|
|
Default value: all-zeroes.
|
|
|
|
[[param-PMP_HARDWIRED_ADDR]]
|
|
===== PMP_HARDWIRED_ADDR
|
|
|
|
Values of pmpaddr registers whose PMP_HARDWIRED bits are set to 1. Has no effect on PMP regions which are not hardwired.
|
|
|
|
Default value: all-zeroes.
|
|
|
|
[[param-PMP_HARDWIRED_CFG]]
|
|
===== PMP_HARDWIRED_CFG
|
|
|
|
Values of pmpcfg registers whose PMP_HARDWIRED bits are set to 1. Has no effect on PMP regions which are not hardwired.
|
|
|
|
Default value: all-zeroes.
|
|
|
|
[[param-DEBUG_SUPPORT]]
|
|
===== DEBUG_SUPPORT
|
|
|
|
Support for run/halt and instruction injection from an external Debug Module,
|
|
support for Debug Mode, and Debug Mode CSRs.
|
|
|
|
Requires: <<param-CSR_M_MANDATORY>>, <<param-CSR_M_TRAP>>.
|
|
|
|
Default value: 0
|
|
|
|
[[param-BREAKPOINT_TRIGGERS]]
|
|
===== BREAKPOINT_TRIGGERS
|
|
|
|
Number of hardware breakpoints. A breakpoint is implemented as a trigger that
|
|
supports only exact execution address matches, ignoring instruction size.
|
|
That is, a trigger which supports type=2 execute=1 (but not store/load=1,
|
|
i.e. not a watchpoint).
|
|
|
|
Requires: <<param-DEBUG_SUPPORT>>
|
|
|
|
Default value: 0
|
|
|
|
==== External interrupt support
|
|
|
|
[[param-NUM_IRQS]]
|
|
===== NUM_IRQS
|
|
|
|
Number of external IRQs implemented in meiea, meipa, meifa and meipra, if
|
|
<<param-CSR_M_TRAP>> is enabled. Minimum 1, maximum 512.
|
|
|
|
Default value: 32
|
|
|
|
[[param-IRQ_PRIORITY_BITS]]
|
|
===== IRQ_PRIORITY_BITS
|
|
|
|
Number of priority bits implemented for each interrupt in meipra. The
|
|
number of distinct levels is (1 << IRQ_PRIORITY_BITS). Minimum 0, max 4.
|
|
Note that having more than 1 priority level with a large number of IRQs
|
|
will have a severe effect on timing.
|
|
|
|
Default value: 0
|
|
|
|
==== Identification Registers
|
|
|
|
[[param-MVENDORID_VAL]]
|
|
===== MVENDORID_VAL
|
|
|
|
Value of the <<reg-mvendorid>> CSR. JEDEC JEP106-compliant vendor ID, or
|
|
all-zeroes. 31:7 is continuation code count, 6:0 is ID. Parity bit is not
|
|
stored.
|
|
|
|
Default value: all-zeroes.
|
|
|
|
[[param-MIMPID_VAL]]
|
|
===== MIMPID_VAL
|
|
|
|
Value of the <<reg-mimpid>> CSR. Implementation ID for this specific version of Hazard3. Should be a git hash, or all-zeroes.
|
|
|
|
Default value: all-zeroes.
|
|
|
|
[[param-MHARTID_VAL]]
|
|
===== MHARTID_VAL
|
|
|
|
Value of the <<reg-mhartid>> CSR. Each Hazard3 core has a single hardware thread. Multiple cores should have unique IDs.
|
|
|
|
Default value: all-zeroes.
|
|
|
|
[[param-MCONFIGPTR_VAL]]
|
|
===== MCONFIGPTR_VAL
|
|
|
|
Value of the <<reg-mconfigptr>> CSR. Pointer to configuration structure blob,
|
|
or all-zeroes. Must be at least 4-byte-aligned.
|
|
|
|
Default value: all-zeroes.
|
|
|
|
==== Performance/size options
|
|
|
|
[[param-REDUCED_BYPASS]]
|
|
===== REDUCED_BYPASS
|
|
|
|
Remove all forwarding paths except X->X (so back-to-back ALU ops can still run
|
|
at 1 CPI), to save area. This has a significant impact on per-clock
|
|
performance, so should only be considered for extremely low-area
|
|
implementations.
|
|
|
|
Default value: 0
|
|
|
|
[[param-MULDIV_UNROLL]]
|
|
===== MULDIV_UNROLL
|
|
|
|
Bits per clock for multiply/divide circuit, if present. Must be a power of 2.
|
|
|
|
Default value: 1
|
|
|
|
[[param-MUL_FAST]]
|
|
===== MUL_FAST
|
|
|
|
Use single-cycle multiply circuit for MUL instructions, retiring to stage 3.
|
|
The sequential multiply/divide circuit is still used for MULH*
|
|
|
|
Default value: 0
|
|
|
|
[[param-MUL_FASTER]]
|
|
===== MUL_FASTER
|
|
|
|
Retire fast multiply results to stage 2 instead of stage 3.
|
|
Throughput is the same, but latency is reduced from 2 cycles to 1 cycle.
|
|
|
|
Requires: <<param-MUL_FAST>>.
|
|
|
|
Default value: 0
|
|
|
|
[[param-MULH_FAST]]
|
|
===== MULH_FAST
|
|
|
|
Extend the fast multiply circuit to also cover MULH*, and remove
|
|
the multiply functionality from the sequential multiply/divide circuit.
|
|
|
|
Requires: <<param-MUL_FAST>>
|
|
|
|
Default value: 0
|
|
|
|
[[param-FAST_BRANCHCMP]]
|
|
===== FAST_BRANCHCMP
|
|
|
|
Instantiate a separate comparator (eq/lt/ltu) for branch comparisons, rather
|
|
than using the ALU. Improves fetch address delay, especially if `Zba`
|
|
extension is enabled. Disabling may save area.
|
|
|
|
Default value: 1
|
|
|
|
[[param-RESET_REGFILE]]
|
|
===== RESET_REGFILE
|
|
|
|
Whether to support reset of the general purpose registers. There are around 1k
|
|
bits in the register file, so the reset can be disabled e.g. to permit
|
|
block-RAM inference on FPGA.
|
|
|
|
Default value: 1
|
|
|
|
[[param-BRANCH_PREDICTOR]]
|
|
===== BRANCH_PREDICTOR
|
|
|
|
Enable branch prediction. The branch predictor consists of a single BTB entry
|
|
which is allocated on a taken backward branch, and cleared on a mispredicted
|
|
nontaken branch, a fence.i or a trap. Successful prediction eliminates the
|
|
1-cyle fetch bubble on a taken branch, usually making tight loops faster.
|
|
|
|
Requires: <<param-EXTENSION_ZIFENCEI>>
|
|
|
|
Default value: 1
|
|
|
|
[[param-MTVEC_WMASK]]
|
|
===== MTVEC_WMASK
|
|
|
|
MTVEC_WMASK: Mask of which bits in mtvec are writable. Full writability (except for bit 1) is
|
|
recommended, because a common idiom in setup code is to set mtvec just
|
|
past code that may trap, as a hardware `try {...} catch` block.
|
|
|
|
|
|
* The vectoring mode can be made fixed by clearing the LSB of MTVEC_WMASK
|
|
* In vectored mode, the vector table must be aligned to its size, rounded
|
|
up to a power of two.
|
|
|
|
Default: All writable except for bit 1.
|
|
|
|
=== Interfaces (Top-level Ports)
|
|
|
|
TODO lol
|